Preference Cluster Mapping
Improvements to Preference Cluster Mapping
Continual consumer testing is a time and cost consuming venture in the development of new products in which several variations are tested before a "winner" is agreed upon. Descriptive testing can offer a cheaper solution but provides no information about consumer acceptance of the product. Partial Least Squares regression models provide a statistical method to relate descriptive profiles collected from a trained panel to consumer acceptance data collected from a large consumer group, thus providing an avenue to shorten the time and reduce the cost of product development.
The use of PLS models eliminates drawback of both Ordinary Least Squares (OLS) Regression and Principal Component Regression (PCR). OLS requires more samples (products) than variables to be included in the model, which is typically not the case when attempting to compare consumer and descriptive data. In fact, it is not unusual to present only 8-10 products to the consumer and descriptive panels yet collect information on a couple dozen descriptive attributes. PCR eliminates this difficulty, as well as the problem of multicollinearity that is present in much of the descriptive data. However, the first principal component, formed from the descriptive data, is not necessarily related to consumer acceptance thereby weakening the model for predictive purposes. The use of PLS through Preference Cluster Mapping eliminates both of these difficulties. Further, because it still fits a model to the data it remains possible to predict consumer liking of new products from the existing model by simply running additional descriptive panels.
One of the drawbacks to Preference Cluster Mapping has been the lack of statistical testing to determine the most parsimonious (simplest) model. Rather, chosen models have tended to include most, if not all of the descriptive attributes and inference has been restricted to the graphical inspection of loading and score plots. From these plots "drivers" of liking, both positive and negative, are identified based on the observed proximity of these descriptive attributes to consumer liking. This often leads to a large number of descriptive attributes that appear to be driving consumer liking and thus makes the identification of an "ideal" product much more complicated both to identify and to create. Nevertheless, the use of Partial Least Square modeling has allowed such "ideal" models to be identified by focusing on the positive and negative drivers and adjusting the descriptive levels of these attributes until such an "ideal" product is identified.
Recent research, now incorporated into The Unscrambler® , eliminates several of the drawbacks discussed above. The implementation of a jackknife procedure provides for the construction of a statistical test to determine the importance of each of the descriptive attributes with respect to the final model. A final model can then be chosen that includes only the statistically important descriptive attributes. This model is much simpler than a model including most if not all of the original variables and when used for prediction the simpler model typically shows a decrease in prediction error. Further, a simpler model makes the identification and creation of an "ideal" product much more direct as fewer variables need to be examined for their effect on consumer acceptance. Development costs are also lowered as descriptive panels only need to collect information on a handful of attributes for use in predicting consumer acceptance of these new products.